Minimizing variance of reaction rate estimation in weighted ensemble simulation using synthetic molecular dynamics

Weighted ensemble (WE) simulation employs multiple trajectories which intermittently are replicated or pruned to enhance the sampling of rare events. WE has recently attracted interest due to its conceptual simplicity, sampling power, toolkit accessibility, versatility for any type of stochastic process, and statistical exactness. However, although WE is unbiased, its performance as measured by run-to-run variance can depend significantly on the two types of hyperparameters: WE bins, which partition the conformation space, and WE allocation, which determines how many trajectories are maintained in each bin. Recent mathematical developments have shed light on how different steps of the WE procedure contribute to the overall variance and suggested parametrization strategies to minimize the variance of flux estimation. However, these results are based on asymptotic behaviors that are not achievable in simulations of realistic molecular systems. In this work, we explore the efficacy of a range of WE optimization strategies using synthetic molecular dynamics, a tractable yet nontrivial proxy to explicit MD based on fine-grained Markov state models.